Anthropic had a week that reads like a venture capitalist’s fever dream. On May 28, it shipped Claude Opus 4.8, a model that lifts agentic coding scores from 64.3% to 69.2% and introduces “dynamic workflows” in Claude Code for tackling problems too large to fit in a single context window. The new model is also, according to Anthropic, more honest about uncertainty—more willing to flag that it’s stuck, less likely to bluff forward. That quiet feature may matter more than the benchmark bump for anyone running long-horizon autonomous tasks.
Ten days earlier, the company announced it had acquired Stainless for more than $300 million. Stainless, a New York startup founded by former Stripe engineer Alex Rattray, automates the creation and maintenance of software development kits—the libraries every developer uses to interact with an API. Its clients included OpenAI and Google. With the acquisition, Anthropic pulled a piece of shared developer infrastructure out of the commons and into its own house, then shut down the hosted product. OpenAI and Google now have to rebuild or migrate. That’s not an accident; it’s a structural move.
Sitting under both stories: Anthropic closed a $30 billion growth round at a valuation above $900 billion—its second $30 billion raise this calendar year. Annualized revenue went from $14 billion in February to $30 billion in April. Whatever Anthropic is selling, enterprises are buying it fast enough to double the run rate in ten weeks.
The acquisition wave
Anthropic wasn’t alone in going shopping. In a five-day span in late May, four frontier labs each absorbed a startup: Mistral bought Vienna’s Emmi AI (physics-aware models for industrial engineering—airflow, heat transfer, material stress); Google DeepMind structured an $80–90 million licensing deal to absorb the entire Contextual AI team, using a structure explicitly designed to avoid antitrust classification as a merger; and Meta acqui-hired the team from Dreamer.
The pattern is legible. The frontier model race is expensive enough that competing on raw capability alone is unsustainable. What labs can do instead is lock up the specific researchers, tools, and domain knowledge that sit between the model layer and the end customer. Stainless sat between Anthropic’s API and every developer who called it. Emmi sits between physics simulation and industrial customers who can’t use a generic transformer for aerodynamics. Contextual AI sits between RAG pipelines and enterprise search. None of these is a bet on one capability; each is a bet on defensibility.
The search revolt
Google’s I/O 2026 conference on May 19 was a show of force: Gemini 3.5 Flash, managed agents in the API, a new omni model, and the most significant overhaul of Google Search in more than 25 years—replacing blue links with AI agents as the default experience.
The response from part of the internet was immediate. DuckDuckGo reported U.S. iOS app installs averaging 33% week-over-week growth the week after I/O, peaking at nearly 70% on a single day. Its AI-free search page, noai.duckduckgo.com, averaged 22.7% weekly growth. Not everyone wants an AI mediating their search results.
This is underrated. Google has spent enormous capital assuming that more AI in search is self-evidently better. The DuckDuckGo data suggests the assumption doesn’t universally hold—when AI is optional and useful, users love it; when it’s mandatory and in the way, some proportion will route around it.
The consolidation push and the search revolt are both, at root, about the same question: who controls the interface between human intent and machine capability. Labs are buying that interface wholesale; some users are walking away from it entirely.